FastHCS: Performs the FastHCS algorithm for robust PCA.

Description

Computes a robust PCA model with q components for an n by p matrix of multivariate data using the FastHCS algorithm.

Usage

FastHCS(x,nSamp=NULL,alpha=0.5,q=10,seed=1)

Arguments

A numeric n (n>5*q) by p (p>1) matrix or data frame.

nSamp

A positive integer giving the number of resamples required; "nsamp" may not be reached if too many of the q-subsamples, chosen out of the observed vectors, are in a hyperplane. If "nSamp" is omitted, it is calculated to provide a breakdown point of "alpha" with probability 0.99.

alpha

Numeric parameter controlling the size of the active subsets i.e., "h=quanf(alpha,n,q)". Allowed values are between 0.5 and 1 and the default is 0.5.

Number of principal components to compute. Note that p>q>1, 1<q<n. Default is q=10.

seed

Starting value for random generator. Default is seed = 1.

Value

A list with components:

rawBest:

The indexes of the h members of H*, the raw FastHCS optimal subset.

obj:

The FastHCS objective function corresponding to H*, the selected subset of h observations.

rawDist:

Outlyingness index of the data on the raw q-dimensonal subset that initialized H*.

best:

the indexes of the members of the H+, the FastHSC subset after the C-steps.

center:

the p-vector of column means of the observations with indexes in best.

loadings:

the (rank q) loadings matrix of the observations with indexes in best.

eigenvalues:

the first q) eigenvalues of the observations with indexes in best.

od:

the orthogonal distances of the centered data wrt to the subspace spanned by the loadings matrix.

sd:

the score distances of the data projected on the subspace spanned by the loadings matrix with respect to the estimated center.

cutoff.od:

the cutoff for the vector of orthogonal distances.

cutoff.sd:

the cutoff for the vector of score distances.

scores

The value of the projected on the space of the principal components data (the centred data multiplied by the loadings matrix) is returned. Hence, cov(scores) is the diagonal matrix diag(eigenvalues).

References

Schmitt E. and Vakili K. and (2015). Robust PCA with FastHCS. (http://arxiv.org/abs/1402.3514)

Examples

Run this code

# NOT RUN {
## testing outlier detection
n<-100
p<-30
Q<-5
set.seed(123)
x0<-matrix(rnorm(n*p),nc=p)
x0[1:30,]<-matrix(rnorm(30*p,4.5,1/100),nc=p)
z<-c(rep(0,30),rep(1,70))
nStarts<-FHCSnumStarts(q=Q,eps=0.4)
Fit<-FastHCS(x=x0,nSamp=nStarts,q=Q)
z[Fit$best]
plot(Fit,col=(!z)+1,pch=16)

## testing outlier detection, different value of alpha
n<-100
p<-30
Q<-5
set.seed(123)
x0<-matrix(rnorm(n*p),nc=p)
x0[1:20,]<-matrix(rnorm(20*p,4.5,1/100),nc=p)
z<-c(rep(0,20),rep(1,80))
nStarts<-FHCSnumStarts(q=Q,eps=0.25)
Fit<-FastHCS(x=x0,nSamp=nStarts,q=Q,alpha=0.75)
z[Fit$best]


#testing exact fit
n<-100
p<-5
Q<-4
set.seed(123)
x0<-matrix(rnorm(n*p),nc=p)
x0[1:30,]<-matrix(rnorm(30*p,4.5,1/100),nc=p)
x0[31:100,4:5]<-x0[31:100,2]
z<-c(rep(0,30),rep(1,70))
nStart<-FHCSnumStarts(q=Q,eps=0.4)
results<-FastHCS(x=x0,nSamp=nStart,q=Q)
z[results$best]
results$obj

#testing rotation equivariance
n<-100
p<-10
Q<-3
set.seed(123)
x0<-scale(matrix(rnorm(n*p),nc=p))
A<-diag(rep(1,p))
A[1:2,1:2]<-c(0,1,-1,0)
x1<-x0%*%A
nStart<-FHCSnumStarts(q=Q,eps=0.4)
r0<-FastHCS(x=x0,nSamp=nStart,q=Q,seed=0)
r1<-FastHCS(x=x1,nSamp=nStart,q=Q,seed=0)
max(abs(log(r1$eigenvalues[1:Q]/r0$eigenvalues[1:Q])))
# }